Query Processing, Resource Management, and Approximation in a Data Stream Management System

1. Query Processing, Resource Management, and Approximationin a Data Stream Management System

2. Introduction • Two fundamental differences between DSMS and DBMS • In addition to managing traditional stored data such as relations, a DSMS must handle multiple continuous, unbounded, possibly rapid and time-varying data streams. • Due to the continuous nature of the data, a DSMS typically supports long-running continuous queries, which are expected to produce answers in a continuous and timely fashion. • Goal – Building general purpose DSMS • supports a declarative query language and can cope with high data rates and thousands of continuous queries • multi-query optimization, judicious resource allocation, and sophisticated scheduling to achieve high performance • environments where data rates and query load may exceed available resources • In these cases the system is designed to provide approximate answers to continuous queries.

3. Query Language • CQL: Continuous Query Language. • All queries are continuous, as opposed to the one-time queries supported by an DBMS • STREAM vs Relation • Streams have the notion of an arrival order, they are unbounded, and they are append-only. A stream can be thought of as a set of <t, s> pairs indicating that a tuple, s, arrives on the stream at time, t. Streams may result from queries or subqueries • Relations are unordered, and they support updates and deletions as well as insertions (all of which are timestamped). In addition to relations stored by the DSMS, relations may result from queries or subqueries • CQL: Fromclause may be followed by an optional sliding window specification, enclosed in brackets, and an optional sampling clause • a window specification consists of an optional partitioning clause (Partition By), a mandatory window size (Rows, Range), and an optional filtering predicate (WEHER) • A sampling clause (Sample) specifies that a random sample of the data elements from the stream

4. Examples Q1: Select Count(*) From Requests S [Range 1 Day Preceding] Where S.domain = ‘stanford.edu’ Q2: Select Count(*) From Requests S [Partition By S.client_id Rows 10 Preceding Where S.domain = ‘stanford.edu’] Where S.URL Like ‘http://cs.stanford.edu/%’ Q3: Select T.URL From (Select client_id, URL From Requests S, Domains R Where S.domain = R.domain And R.type = ’commerce’) T Sample(10) Where T.client_id Between 1 And 1000

5. Formal Semantics • A window specification is applied to a stream up to a specific time t and the result is a finite set of tupleswhich is treated as a relation. • Istream(for “insert stream”) applied to relation R contains a stream element <t, s> whenever tuple s is in R at time t but not in at time t-1 • Dstream (for “delete stream”) applied to relation R contains a stream element <t, s> whenever tuple s is in at time t-1 but not in at time t • Using the mappings in the figure, CQL queries can freely mix relations and streams. • The result of a query at time t is obtained by taking all relations at time t, all streams up to time t converted to relations by their window specifications, and applying conventional relational semantics. If the outermost operator is Istream or Dstream then the query result is converted to a stream, otherwise it remains as a relation.

6. Stream Ordering and Timestamps • Assumptions • So that we can evaluate row-based (Rows) and timebased (Range) sliding windows, all stream elements arrive in order, timestamped according to a global clock. • So that we can coordinate streams with relation states, all relation updates are timestampedaccording to the same global clock as streams. • So that we can generate query results, the global clock provides periodic “heartbeats” that tell us when no further stream elements or relation updates will occur with a timestamp lower than the heartbeat value.

7. Inactive and Weighted Queries • Inactive query • When a query is inactive, the system may not maintain the answer to the query as new data arrives. • an inactive query may be activated at any time • Influences decisions about query plans and resource allocation • Weight • Indicating relative importance of a query • Weights might also influence scheduling decisions • inactive queries may be thought of as queries with negligible weight

8. Query Plans • Three different types of components of a query plan • Query operators, similar to a traditional DBMS. Each operator reads a stream of tuples from a set of input queues, processes the tuples based on its semantics, and writes its output tuples into a single output queue. • Inter-operator queues, also similar to the approach taken by some traditional DBMS’s. Queues connect different operators and define the paths along which tuplesflow as they are being processed. • Synopses, used to maintain state associated with operators

9. Synopsis • A synopsis summarizes the tuples seen so far at some intermediate operator in a running query plan, as needed for future evaluation of that operator. • For full precision a join operator must remember all the tuples it has seen so far on each of its input streams, so it maintains one synopsis for each • Synopsis sizes grow without bound if full precision is expected in the query result • An important feature to support is synopses that use some kind of summarization technique to limit their size • fixed-size hash tables, sliding windows, reservoir samples, quantile estimates, and histograms.

10. Generic methods of Operator and Synopsis • Operator Class • create, with parameters specifying the input queues, output queue, and initial memory allocation. • changeMem, with a parameter indicating a dynamic decrease or increase in allocated memory. • run, with a parameter indicating how much work the operator should perform before returning control to the scheduler • Synopsis Class • create, with a parameter specifying an initial memory allocation. • changeMem, with a parameter indicating a dynamic decrease or increase in allocated memory. • insert and delete, with a parameter indicating the data element to be inserted into or deleted from the synopsis. • query, whose parameters and behavior depend on the synopsis type. For example, in a hash-table synopsis this method might look for matching tuples with a particular key value, while for a sliding window synopsis this method might support a full window scan.

11. An Example of query plans for Q1 and Q2 • Two plans contain three operators • Two plans contain four synopses • Two plans contain four queues • Two plans share a subplan joining streams R and S • Execution of query operators is controlled by a global scheduler

12. Resource Sharing in Query Plans • Important topics are yet to be addressed • For now we are considering resource sharing and approximation separately. That is, we do not introduce sharing that intrinsically introduces approximate query results, such as merging subexpressions with different window sizes, sampling rates, or filters. Doing so may be a very effective technique when resources are limited, but we have not yet explored it in sufficient depth to report here. • Our techniques so far are based on exact common subexpressions. Detecting and exploiting subexpression containment is a topic of future work that poses some novel challenges due to window specifications, timestamps and ordering, and sampling in our query language. • A shared queue maintains a pointer to the first unread tuplefor each operator that reads from the queue, and it discards tuplesonce they have been read by all parent operators. • If two queries with a common subexpressionproduce parent operators with very different consumption rates, then it may be preferable not to use a shared subplan.

13. Synopsis Sharing • When several operators read from the same queue, and when more than one of those operators builds some kind of synopsis, then it may be beneficial to introduce synopsis sharing • Which operator is responsible for managing the shared synopsis (e.g., allocating memory, inserting tuples)? • If the synopses required by the different operators are not of identical types or sizes, is there a theory of “synopsis subsumption” (and synopsis overlap) that we can rely on? • If the synopses are identical, how do we cope with the different rates at which operators may “consume” data in the synopses?

14. Resource Management • Relevant resources in a DSMS • memory, computation, I/O if disk is used, and network bandwidth • Focusing initially on memory consumed by query plan synopses and queues • in many cases reducing memory overhead has a natural side-effect of reducing other resource requirements as well • Resource Management • An algorithm for incorporating known constraints on input data streams to reduce synopsis sizes • An algorithm for operator scheduling that minimizes queue sizes

15. Exploiting Constraints Over Data Streams • Additional information about streams • gathering statistics over time • constraint specifications at stream-registration time • Use this information to reduce resource requirements without sacrificing query result precision • An alternate and more dynamic technique is for the streams to contain punctuations, which specify run-time constraints • Example [Join Order(order_id) and Fulfillments(item_id)] • In the general case, requires synopses of unbounded size • What if we know that all tuples for a given orderID and itemID arrive on O before the corresponding tuples arrive on F? • We need not maintain a join synopsis for the F operand at all • In practice, constraints may not be adhered to by data streams strictly • no more than k tupleswith a different orderID appear between two tuples with same orderID

16. Scheduling (1) • Query plans are executed via a global scheduler, which calls the run methods of query plan operators • Example • O1 operates on input queue q1 and writing results to queue q2which is the input queue of O2 • O1 takes one time unit to operate on a batch of n tuples from q1 and it has 20% selectivity • O2 takes one time unit to operate on tuples and let us assume that its output is not queued by the system (final result)

17. Scheduling (2) • Two possible scheduling strategies • Tuples are processed to completion in the order they arrive at q1 Each batch of n tuples in q1 is processed by O1 and then O2 based on arrival time, requiring two time units overall. • If there is a batch of n tuples in q1, then O1 operates on them using one time unit, producing n/5 new tuples in q2. Otherwise, if there are any tuplesin q2, then up to n/5 of these tuples are operated on by Q2, requiring one time unit.

18. Approximations • The system will not be able to provide continuous and timely exact answers to all registered queries • multiple unbounded and possibly rapid incoming data streams • multiple complex continuous queries with timeliness requirements • finite computation and memory resources

19. Static Approximation • Window Reduction • two considerations • If W is a duplicate-elimination operator, then shrinking W’s window can actually increase its output rate • If W is part of the right-hand subtree of a negation construct (e.g., NOT EXISTS or EXCEPT), then reducing the size of output may have the effect of increasing output further up the query plan • Sampling Rate Reduction • Although changing the sampling rate at an operator will not reduce the resource requirements of it will reduce the output rate.

20. Dynamic Approximation • Synopsis Compression • maintaining a sample of the intended synopsis content • using histograms for aggregation • compressed wavelets for aggregation • using Bloom filters for duplicate elimination, set difference, or set intersection • Sampling and Load Shedding • approximation techniques that reduce queue sizes • introduce one or more sample operators into the query plan, or to reduce the sampling rate at existing operators • simply drop tuples from queues when the queues grow too large, a technique sometimes referred to as load shedding • load shedding may drop chunks of tuples at a time

21. Resource Management and Approximation: Discussion • Many important challenges • We need a means of monitoring synopsis and queue sizes and determining when dynamic reduction measures (e.g., window size reduction, load shedding) should kick in • Even if we have a good algorithm for initial allocation of memory to synopses and queues, we need a reallocation algorithm to handle the inevitable changes in data rates and distributions • The ability to add, delete, activate, and deactivate queries at any time forces all resource allocation schemes, including static ones, to provide a means of making incremental changes